R[AI]DAR

Category intelligence

Social Media Briefing

All 150 current items, analyzed and ranked.

Executive synthesis

Social Media Summary

Executive Signal

  • Open-process frontier development compresses: Scaled multimodal models, open training recipes, and local-serving economics reshape competition, while benchmark credibility and AI governance discourse intensify.

Priority Developments

  • Multimodal frontier expands across vision and touch: DeepSeek-V4-Flash-Vision-Exp approaches Opus-4.8 on agent benchmarks while T-Rex co-designs tactile sensing with a 50-hour open dataset.
  • Open-process frontier training scales: Marin 535B-A23B begins a ~3-month run on 18.75T tokens across 11x GB200 NVL72, gated by a four-rung scaling ladder for forecasting.
  • Frontier model economics compress: OpenAI cuts GPT-5.6 Sol API pricing >20% for three months; Stoica argues local consumer-GPU serving reshapes AI unit economics.
  • Benchmark credibility and AI ethics face scrutiny: Chollet critiques NVIDIA's 100% ARC-AGI-3 demo claim; Gebru challenges TESCREAL-linked research as pseudo-scientific racism.
  • Infrastructure precision tightens: vLLM SkyRL IsoExec achieves near-bitwise parity between rollout and trainer engines for Qwen3.5-35B-A3B on DAPO.

Leadership Implications

Key Themes

Model Releases & Multimodal Capability · 5Open-Source Pretraining & Robotics Research · 2Benchmark & Agent Capability Claims · 3Reinforcement learning training infrastructure · 1New model announcements · 4Local inference and open-weight economics · 1AI ethics and societal critique · 2Open ML research and training analysis · 3AI Evaluation and Benchmarks · 3AI Ethics and Governance · 1

Primary evidence

All Ranked Signals

Showing 10 of 150 signals 2 remaining
Social Twitter Aug 21

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multim...

By @deepseek_ai

92 score
AI Analysis

Continuing our coverage from yesterday, DeepSeek officially launches DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model on their API platform, claiming text parity with V4-Flash and major gains on multimodal agent benchmarks approaching Opus-4.8. DeepSeek Harness 0.1.1 released with out-of-the-box support.

Source details
DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge. 🔹 On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8. Try it with model='deepseek-v4-flash-vision-exp'. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model. 1/n
model-releasemultimodalopen-weight-modelsDeepSeek
Social Twitter Aug 21

🚢 Marin 535B-A23B started training this week! As usual, the whole process is open. Voyage plan: pret...

By @percyliang

90 score
AI Analysis

Building on yesterday's Social buzz, Percy Liang announces Marin 535B-A23B has begun training, fully open process. Plans 80% pretraining + 20% midtraining on 18.75T tokens across 11x GB200 NVL72 for ~3 months (~2.7e24 FLOPs), preceded by a 4-rung scaling ladder (1.6B to 27.7B) for forecasting.

Source details
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open. Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow. Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
open-sourcepretrainingmixture-of-expertsMoEscaling-lawsMarin
Social Twitter Aug 21

The sense of touch is the most criminally under-explored modality in robotics. Imagine doing sleight...

By @DrJimFan

84 score
AI Analysis

Jim Fan introduces T-Rex, an open methodology co-designing touch as a first-class modality in robotics. Uses an asynchronous mixture-of-transformer with a slow visuomotor expert and fast tactile expert at 4 touch ticks per vision tick, plus a 50-hour tactile robot dataset (largest publicly released).

Source details
The sense of touch is the most criminally under-explored modality in robotics. Imagine doing sleight of hand wearing thick oven mitts. That's exactly how a robot feels today if it were alive. A magnetic piece snapping into place, a paper cup peeling out of a stack, a USB negotiating its way into the port - all invisible to the camera. Learning how to feel must be a full-stack co-designed effort. We are open-sourcing a principled methodology called "T-Rex": 1. Tactile as first-class citizen of
roboticstactile-sensingopen-sourcemultimodaldata
Social Twitter Aug 21

As we continue to push the frontier of capabilities while improving efficiency, we're dropping API a...

By @OpenAI

82 score
AI Analysis

OpenAI announces over 20% API and credit pricing reduction for GPT-5.6 Sol for the next three months, framing it as part of improving efficiency alongside capability gains.

Source details
As we continue to push the frontier of capabilities while improving efficiency, we're dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months. t.co/UoTb3hcB2t
pricingOpenAIAPI-economics
Social Twitter Aug 21

This is very nice work from NVIDIA. Like all high-performing approaches on ARC-AGI-3, it uses deep l...

By @fchollet

82 score
AI Analysis

Following yesterday's Social coverage, François Chollet — creator of ARC-AGI — critiques NVIDIA's 100% ARC-AGI-3 claim, clarifying that the public demonstration set is only a subset of the full benchmark and equating perfect demo scores to clearing a tutorial level.

Source details
This is very nice work from NVIDIA. Like all high-performing approaches on ARC-AGI-3, it uses deep learning-guided on-the-fly synthesis of symbolic world models, i.e. navigating the world by generating programs to represent what you know. To be clear, like with several other recent claims, scoring 100% on the public demonstration set is not the same as "scoring 100% on the ARC-AGI-3 benchmark". It would be like saying you beat a videogame because you cleared the tutorial level.
benchmarksARC-AGIcritiquemethodologyevaluation
Social Twitter Aug 21

I KNEW this was gonna get connected back to the TESCREAL bozos. Longtermism is the L in TESCREAL. ...

By @timnitGebru

78 score
AI Analysis

Timnit Gebru critiques a paper connected to the TESCREAL movement (transhumanism, eugenics, singularity, rationalism, effective altruism, longtermism), calling out longtermism and pseudo-scientific 'dysgenic' claims as thinly veiled racism in AI-adjacent research.

Source details
I KNEW this was gonna get connected back to the TESCREAL bozos. Longtermism is the L in TESCREAL. This "paper" is by the guy who hired Cofnas at Ghent. Reminder "dysgenic" is the opposite of "eugenic," they mean Black people and other undesirables reproducing too much. t.co/N6oIbEVK8c
AI ethicseugenics criticismTESCREAL/longtermismAI governance
Social Twitter Aug 21

In RL training, a vLLM rollout engine and a Megatron trainer can run the same policy yet disagree on...

By @vllm_project

78 score
AI Analysis

Detailed technical post from the vLLM project explaining SkyRL's IsoExec component, which addresses floating-point non-associativity between rollout engines and trainers (e.g., Megatron) by aligning rounding-sensitive execution across tensor, expert, and sequence-parallel layouts, achieving near-bitwise parity for Qwen3.5-35B-A3B on DAPO with 8xH100.

Source details
In RL training, a vLLM rollout engine and a Megatron trainer can run the same policy yet disagree on a token's logprob due to floating-point non-associativity. SkyRL's IsoExec combines an execution contract with a unified model, aligning rounding-sensitive execution choices across rollout and training. Bitwise parity holds across different TP, EP, and SP layouts. For Gated DeltaNet, the chunkwise-parallel recurrent algorithm makes parallel training and prefill bitwise identical to recurrent de
reinforcement learninginfrastructurevLLMmodel trainingnumerical precision
Social Twitter Aug 21

The ability to serve frontier-scale models locally fundamentally shifts the economics of AI. By opti...

By @istoica05

76 score
AI Analysis

Ion Stoica argues that serving frontier-scale models locally on consumer GPUs fundamentally shifts AI economics by lowering the barrier to entry for agentic workflows.

Source details
The ability to serve frontier-scale models locally fundamentally shifts the economics of AI. By optimizing memory and compute to fit these models on consumer GPUs, the barrier to entry for building and deploying robust agentic workflows is significantly lowered.
local inferenceconsumer GPUsAI economicsagentic workflows
Social Twitter Aug 21

Our general-purpose coding agent just scored 100% on the ARC-AGI-3 interactive reasoning benchmark. ...

By @NVIDIAAI

73 score
AI Analysis

Continuing our coverage from yesterday, NVIDIA claims its general-purpose coding agent AVO scored 100% on the ARC-AGI-3 interactive reasoning benchmark across all 183 levels in 25 public environments without instructions or rules.

Source details
Our general-purpose coding agent just scored 100% on the ARC-AGI-3 interactive reasoning benchmark. NVIDIA AVO completed all 183 levels across all 25 public environments, figuring out what to do with no instructions, explicit rules, or stated goals. t.co/UgROuDrMtn
benchmarksagentsNVIDIAARC-AGI
Social Twitter Aug 21

🚨👇 Republicans are abandoning data centers faster than rats abandon sinking ships. Perhaps no indus...

By @GaryMarcus

72 score
AI Analysis

Continuing our coverage from yesterday, DeepSeek announces multimodal API support via deepseek-v4-flash-vision-exp, with image tokenization billing (up to 384 tokens per image), compatibility with Chat Completions/Messages/Responses endpoints, and support for mixed text and image inputs.

Source details
🚨👇 Republicans are abandoning data centers faster than rats abandon sinking ships. Perhaps no industry in history has spoiled its own prospects faster than the AI industry. Through a mixture of greed, stupidity, and arrogance, the heroes of 2023 have become the villains of 2026.
DeepSeekmultimodal AIAPI updatesvision models